NVIDIA PAIR Is a LAN Scheduler for Local Python Agents

Python people have been gluing GPUs together with SSH, tmux, and a homegrown queue for years. At IFA 2026, NVIDIA shipped a named version of that glue. Personal AI Router, or PAIR, is free, open source, and meant to notice idle machines on your LAN and send inference there. It talks to Ollama and LM Studio. It has a GUI and a terminal. The interesting part for this site is not the booth. It is that your next automation job may be an agent with a skill file, not a cron calling subprocess.

We already have a production-agents piece and a LangChain / CrewAI tour. Those are library choices. PAIR is plumbing. Treat it like Celery for tokens.

What NVIDIA actually announced

Gerardo Delgado’s September 3 post lists four things. Simplified local GPU support is coming to Hermes Agent, OpenClaw, and Perplexity’s Portable Computer. llama.cpp and vLLM optimizations claim up to 1.9x faster local inference, also exposed through LM Studio and Ollama. PAIR is the router. RTX Spark Windows PCs are an October hardware story; skip them unless you were already buying a box.

PAIR’s job is narrow. Agentic workflows split work into parallel jobs. Those jobs usually pile onto one GPU. PAIR discovers compatible PCs on the local network and sends independent inference requests to whoever has capacity. Machines can join and leave. SMBtech’s write-up repeats the household stat NVIDIA is using: more than half of U.S. homes have two or more PCs, and a lot of that silicon sits idle. The beta runs on Windows, macOS, and Linux. Hardware floor: GeForce RTX 20-series and newer, RTX PRO Turing and newer, DGX Spark, Apple M4 and newer.

That last item is the compatibility tell. This is not an NVIDIA-only toy. If your second machine is an M4 Mac Mini running Ollama, it is in the pool. If your second machine is a 2018 office PC with no discrete GPU, it is not.

The demo NVIDIA likes is a Hermes “Sunday Reset”: sort a cluttered inbox, decide what is now, later, or skip. Hermes splits that into subagents. PAIR spreads the subagents. Your main PC can keep compiling, or gaming, while the inbox job runs on the box under the desk. That is a scheduler. Write it down that way in the runbook, not as “AI magic.”

Speed claims, with the caveats attached

llama.cpp: up to 1.9x throughput on a GeForce RTX 5090 from kernel work, speculative decoding, and faster prefill. vLLM: 1.2x on an RTX PRO 6000 Blackwell Workstation Edition, up to 1.4x on two DGX Spark machines clustered, with new XQA attention kernels in FlashInfer. Those numbers are NVIDIA’s. They are available in the backends and through LM Studio and Ollama. They are not a promise about your 3070.

If you already pin model files and measure tokens per second, rerun the same prompt before and after the IFA backend bumps. If you do not measure, you will not know whether PAIR helped or whether you just added a network hop. A LAN router in front of a 7B model on a 2060 can be slower than running the 7B on the 5090 in front of you. PAIR only wins when the jobs are actually independent. A four-step chain that needs the previous hidden state will not fan out.

Hermes, from Nous Research, is getting a Windows one-click path: detect the NVIDIA GPU, pick a model and config, run through bundled llama.cpp with the new kernels. Linux is “expected to follow.” Once it is up, NVIDIA’s description is the usual agent list: tools, context, memory across sessions, reusable skills. Data stays on the GPU box if the model is local. That “if” is doing work. A skill that calls a cloud API is not local, even if PAIR routed the first token.

One-click is a gift and a trap. It removes the afternoon you would have spent installing CUDA-aware llama.cpp. It also hides which binary is serving the model. If you automate anything that touches private data, spend the extra hour confirming the process list: llama.cpp or vLLM, bound to localhost or to 0.0.0.0, auth token on or off. PAIR’s terminal mode is useful here because you can see backends as a list instead of a happy animation.

OpenClaw is on the same “simplified local AI” slide as Hermes and Portable Computer. If your team was already evaluating OpenClaw, wait for the IFA integration notes and then read the Recorded Future section below before you clone a random starter repo. The product category is moving faster than the review process.

How this maps onto a Python automation box

You do not have to adopt Hermes. You can keep a FastAPI worker, a Dramatiq queue, or a cron, and point the heavy calls at Ollama on localhost. PAIR sits under Ollama. Your Python does not change, except the Ollama host might now be http://pair-router:11434 instead of 127.0.0.1.

That is the conservative integration. Test it like any other queue.

1. Pin the model name in config, not in a skill markdown file someone pasted from Discord.
2. Health-check the router before you enqueue. If PAIR has zero backends, fail closed. Do not block the cron for ten minutes.
3. Treat each inference as idempotent. If a PC drops off the LAN mid-job, retry once on another backend. If the job is “delete these emails,” do not retry blindly.
4. Log which hostname ran the job. You will need that the first time a result looks wrong and you do not know which GPU quantized the model.

Network trust is the other operational detail NVIDIA’s post does not dwell on. PAIR discovers machines on the local network. That is great in a house. It is a different conversation in an office VLAN that still has guest Wi-Fi bridged to it. Put PAIR on a tagged subnet that only the GPU boxes can see. Do not advertise inference on the same SSID your visitors use. The terminal interface is the one you want in that setup. A GUI that “just finds PCs” will find PCs you did not mean to include.

Version-pin Ollama models the way you pin Python deps. llama3.1:8b today is not llama3.1:8b in November if the tag moved. PAIR routing a job to a machine that pulled a newer quant is a Heisenbug: same prompt, different box, different answer. Write the digest into the job record.

The aggressive integration is letting Hermes own the workflow and writing skills that shell out to your existing Python. That is convenient. It is also how you accidentally give an LLM subprocess on a machine that can see your mail. Keep the dangerous bits in a Python service with a narrow API. Let the agent call POST /inbox/plan. Do not let it call shutil.rmtree.

If you have been using MCP servers or the loop-engineering agent pattern, PAIR is a new transport, not a new brain. The brain is still your tool policy.

The part of this week’s news that is not a product launch

NVIDIA’s list of local-agent homes includes OpenClaw. Recorded Future’s H1 2026 malware trends note, dated September 3, is the cold shower. In February, VirusTotal reported malicious OpenClaw skills dressed up as useful automation for a local agent ecosystem. Malwarebytes reported fake OpenClaw installers on GitHub, surfaced through search, dropping infostealers and proxy malware.

That is the same distribution path as “helpful Python scripts” on Gist, except the file now has a friendlier name: skill. If PAIR and Hermes make one-click local agents normal, fake skills become a normal payload. Recorded Future’s hardening list is the unfashionable one: audit GitHub and CI credentials, least privilege for package publishing, watch lifecycle scripts, watch new repos appearing under your org. We covered the broader Python supply-chain problem already. This week just named the agent-shaped variant.

Practical rules that fit in a team doc:

• Install PAIR and Ollama from the project’s own release page, not from a Google result titled “easy OpenClaw.”
• Skills are code. Read them. If a skill fetches a URL and execs it, delete it.
• Separate the GPU box from the box that holds mail passwords. PAIR routing across the LAN is useful. Pairing it with your password manager is not.
• CI that publishes skills should use the same GitHub Actions discipline you use for PyPI: locked deps, no curl | bash in a workflow.

A setup that will still make sense in October

Do not wait for Spark. If you have two RTX machines, or an RTX box and an M4, install the PAIR beta on both, point Ollama at a model you already trust, and run one embarrassingly parallel job: ten independent summaries. Measure wall time versus a single GPU. If it scales, keep it. If the router adds 200ms of discovery and you only ever run one job, uninstall it.

Write the test down. Same ten prompts, same temperature, batch on one GPU, then batch through PAIR. Record tokens per second, failure count, and which host ran which item. If host B is always slower, take it out of the pool instead of “load balancing” into a 2060. Idle silicon is only free if it is not making the p95 worse.

Keep a kill switch. PAIR is a network service. If you cannot name the process and the port, you cannot shut it off when a skill goes weird. systemctl stop or the Windows equivalent should be in the same README as the install command.

Keep your real automation in Python. Reports, file watchers, cloud APIs, the scheduled report pipeline — those do not need an agent and they should not wait on a LAN discovery step. Agents help when the next step is fuzzy. PAIR helps when the fuzzy steps are many and independent. Mixing all three in one process is how Sunday Reset becomes Sunday outage.

October will bring more Windows SKUs. The IFA leftover that matters in a repo is smaller: a LAN-aware inference router, faster llama.cpp/vLLM builds in the tools you already run, and a reminder that “skill” is not a trusted file type. Wire the router. Read the skill. Keep exec out of the queue.